[SOLVED] Cs-349 homework 4: reinforcement learning and ethics
5.0
1 customer review
Digital download
Digital download
$25.00
- Availability
- In stock
- Checkout
- One item
Need a hand?
Message us on WhatsApp for payment or download support.
and frozen lake environments. You may discuss the homework with other groups, but do not take any
written record from these discussions. Also, do not copy any source code from the Web. Your
submission must be your own.1. Coding (4.0 points): Your task is to implement two reinforcement learning algorithms:
Multi-armed Bandits (in src/multi_armed_bandits.py)
Q-Learning (in src/q_learning.py)Note that while these reinforcement learning methods inherently depend on randomization, we
provide a src/random.py package that will randomize things in the same way for all students.Please use src.random anywhere that you might have otherwise used np.random. Your goal is
to pass the test suite (contained in tests/). You can test your code by running pytest
tests/test_bandit.py and pytest tests/test_q_learning.py. Once the tests are
passed, you will use your code to answer FRQ 2 and 3.2. Bandits vs. Q-Learning (2.0 points):
a. Run python -m q2.py; it will create three plots: 2a_SlotMachines_Comparison.png,
2a_FrozenLake_Comparison.png, and
2a_SlipperyFrozenLake_Comparison.png. Please read about the FrozenLake
environment (https://gymnasium.farama.org/environments/toy_text/frozen_lake/) . Each plot
will show a comparison of your MultiArmedBandit and QLearning models on the named
environment (e.g., SlotMachines). Include those plots here. For each plot, provide a onesentence description of the most notable trend. Pay attention to the scale on the y-axis.b. In which of the above plots does Q-Learning appear to receive higher rewards on average
than MultiArmedBandit? Provide an explanation for why that happens, based on your
understanding of Q-Learning.c. Following b.: in the environment(s) where MultiArmedBandit was the worse model, is there
any way you could change your choice of hyperparameters so that MultiArmedBandit would
perform as well as Q-Learning? Why or why not?d. In which of the above plots does MultiArmedBandit appear to receive higher rewards on
average than Q-Learning? Provide an explanation for why that happens, based on your
understanding of MultiArmedBandit.e. Following d.: in the environment(s) where Q-Learning was the worse model, is there any
way you could change your choice of hyperparameters so that QLearning would perform as
well as MultiArmedBandit? Why or why not?3. Exploration vs. Exploitation (2.0 points):
a. Look at the code in q3.py and run python -m q3.py and include the plot it creates
(free_response/3a_g0.9_a0.2.png) as your answer to this part. In your own words,
describe what this code is doing.b. Using the above plot, describe what you notice. What seems to be the “best” value of
epsilon? What explains this result?c. The above plot trains agents for 50,000 timesteps each. Suppose we instead trained them for
500,000 or 5,000,000 timesteps. How would you expect the trends to change or remain the
same for each of the three values of epsilon? Give a one-sentence explanation for each value.d. When people use reinforcement learning in practice, it can be difficult to choose epsilon and
other hyperparameters. Instead of trying three options like we did above, suppose we tried 30
or 300 different choices. What might be the danger of choosing epsilon this way if we
wanted to use our agent in a new domain?4. Tic-Tac-Toe (2.0 points): Suppose we want to train a Reinforcement Learning agent to play the
game of Tic-Tac-Toe (see https://en.wikipedia.org/wiki/Tic-tac-toe), and need to construct an
environment with states and actions. Assume our agent will simply choose actions based on the
current state of the game, rather than trying to guess what the opponent will do next.a. What should be the states and actions within the Tic-Tac-Toe Reinforcement Learning
environment? Don’t try to list them all, just describe how the rules of the game define what
states and actions are possible. How does the current state of the game affect the actions you
can take?b. Design a reward function for teaching a Reinforcement Learning agent to play optimally in
the Tic-Tac-Toe environment. Your reward function should specify a reward value for each
of the 3 possible ways that a game can end (win, loss, or draw) as well as a single reward
value for actions that do not result in the end of the game (e.g., your starting move). Explain
your choices.c. Suppose you were playing a more complicated game with a larger board, and you want the
agent to learn to win as fast as possible. How might you change your reward function to
encourage speed?5. Fair ML in the Real World (3.0 points): Read Joy Buolamwini and Timnit Gebru, 2018. “Gender
shades: Intersectional accuracy disparities in commercial gender classification.” Conference on
fairness, accountability and transparency, then use it to help answer the following questions (see
https://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf).a. Buolamwini and Gebru use PPV and FPR as metrics to measure fairness. Find the definition
of these in the paper, then look up the corresponding definition for NPV and FNR (these
appear in the slides). Assuming you were applying for a loan and you know a ML classifier
is deciding whether to grant it to you: would you rather have that decision made by a system
with a high FPR or a high FNR? Why? Provide a detailed justification.b. Assuming you were applying for a loan and you know a ML classifier is deciding whether to
grant it to you: would you rather have that decision made by a system with a high PPV or a
high NPV? Why? Provide a detailed justification.c. What recommendations do Buolamwini and Gebru make regarding accountability and
transparency of ML systems? How does this relate to specific metrics such as PPV or FPR?d. What is intersectional about the analysis conducted by the authors? What does that analysis
show?e. In Section 4.7, the authors say that their “findings … do not appear to be confounded by the
quality of sensor readings.” What do they mean by “confounded” in this context? Why is it
important to the thesis of this paper to check whether their findings are confounded?Submission Instructions
Turn in your homework as a single zip file, in Canvas. Specifically:
1. Create a single pdf file hw4.pdf with the answers to the questions above and summaries of
your results.
2. Create a single ZIP file containing:
o hw4.pdf
o All of your .py code files
3. Turn the zip file in under Homework #4 in Canvas.
Good luck, and have fun!